🔬 TTT-Discover

Learning to Discover at Test Time

Mert Yuksekgonul*, Daniel Koceja*, Xinhao Li*, Federico Bianchi*
Jed McCaleb, Xiaolong Wang, Jan Kautz, Yejin Choi, James Zou†, Carlos Guestrin†, Yu Sun*

Stanford · NVIDIA · Astera Institute · UC San Diego · Together AI

---

TTT-Discover performs reinforcement learning at test time, allowing the LLM to continue training with experience specific to the problem at hand. We achieve new state-of-the-art across mathematics, GPU kernels, algorithms, and biology.

Installation

pip install ttt-discover

Or from source:

pip install -e .

Set environment variables:

export HF_TOKEN="..."
export TINKER_API_KEY="..."      
export WANDB_API_KEY="..."       
export WANDB_ENTITY="..."        

Making your own Environment

To use TTT-Discover for your own application, you should create a new environment. Here are the general steps to make your own environment.

1. Create a new environment that inherits ttt_discover.Environment.

2. Define a reward evaluator that inherits ttt_discover.BaseRewardEvaluator. Optionally, you can use ttt_discover.SandboxRewardEvaluator to run generated code in sandboxes.

3. (Optional) Add initial state definition to your environment.

4. Define a config and run with ttt_discover.discover!

Here is a sample skeleton for a new environment.

# Import requred ttt_discover objects
from ttt_discover import Environment, BaseRewardEvaluator, State, DiscoverConfig, discover


# Define your reward function
class YourReward(BaseRewardEvaluator):

    def get_reward(self, code: str, state: State) -> float:
        # ...add logic here for computing reward

        return {
            "reward": reward,
            "correctness": 1.0,
            "raw_score": raw_score,
            "msg": f"Success; raw_score={raw_score}",
            "result_construction": [], # Could reuse
            "stdout": "", # No stdout
        }


class YourEnv(Environment):
    reward_function = YourReward
    state_type = State # You may define your own state if you wish

    def get_question(self) -> str:
        state_ctx = self.initial_state.to_prompt(100, metric_name="performance")

        return f"""You are an expert mathematician specializing in combinatorial problems and computational geometry. Your task is to ... {state_ctx}."""


config = DiscoverConfig(
    env_type=YourEnv,
    experiment_name="test-run",
    wandb_project="",
)

# Run discovery
discover(config)

Check examples/circle_packing for a fully implemented example.

Key Results

	Mathematics Erdős Overlap ↓	Kernel A100 TriMul ↓	Kernel H100 TriMul ↓	Algorithms AtCoder ↑	Biology Denoising ↑
Best Human	0.380927	4531 μs	1371 μs	566,997	0.64
Prev. Best AI	0.380924	—	—	558,026	—
TTT-Discover	0.380876	2198 μs	1161 μs	567,062	0.71

Domains

Mathematics — Classic open problems in combinatorics and analysis

Task	Erdős Min. Overlap ↓	Autocorr. (AC1) ↓	Autocorr. (AC2) ↑
Best Human	0.380927	1.50973	0.9015
Prev. Best AI	0.380924	1.50314	0.9610
TTT-Discover	0.380876	1.50287	0.9591

Kernel Engineering — GPUMode TriMul competition for triangular matrix multiplication

Task	A100 ↓	H100 ↓	B200 ↓	MI300x ↓
Best Human	4531 μs	1371 μs	1005 μs	2462 μs
TTT-Discover	2198 μs	1161 μs	905 μs	1596 μs

Algorithm Engineering — AtCoder Heuristic Contests on real-world optimization [AHC39] [AHC58]

Task	AHC39 (Geometry) ↑	AHC58 (Scheduling) ↑
Best Human	566,997	847,674,723
Prev. Best AI	558,026	848,373,282
TTT-Discover	567,062	848,414,228

Biology — Single-cell RNA-seq denoising on OpenProblems benchmark

Task	PBMC ↑	Tabula ↑
Best Human	0.64	0.64
TTT-Discover	0.71	0.73

The environments to reproduce results from our paper are under examples/. To run these, please see reproducing.md

Submitit

We provide submitit script to launch ttt-discover as a slurm job across multiple nodes with ray. See submitit_launch.sh for an example.

Security Notice

It is recommended to run all jobs on an isolated network or VPN if using ray. Ray has minimal built-in security protections and should not be exposed on a public or shared network.

Acknowledgments

This work builds on several outstanding projects and communities:

• GPU Mode — Community for GPU kernel optimization and the TriMul competition
• ALE-Bench — AtCoder-based benchmark for LLM evaluation
• Tinker — LLM training recipes and RL framework

Citation

@article{ttt-discover2026,
  title   = {Learning to Discover at Test Time},
  author  = {Yuksekgonul, Mert and Koceja, Daniel and Li, Xinhao 
             and Bianchi, Federico and McCaleb, Jed and Wang, Xiaolong 
             and Kautz, Jan and Choi, Yejin and Zou, James 
             and Guestrin, Carlos and Sun, Yu},
  journal = {arXiv preprint arXiv:2601.16175},
  year    = {2026}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.